To address the safety, efficiency, and adaptability challenges of autonomous path planning for lunar rovers in complex and unknown terrain, this study proposes a path planning method that integrates deep reinforcement learning with Hybrid A*. Multi-level obstacle maps are generated from lunar digital elevation models and local terrain reconstruction, and an environment model integrating terrain features and rover states is constructed. A Deep Deterministic Policy Gradient framework is introduced, including a state encoder, a policy network, and dual Q-networks, to dynamically adapt the search parameters of Hybrid A* and optimize the path-cost function online. Simulation experiments are conducted under simple, moderate, and complex scenarios. The results show that the proposed method achieves a 100% path planning success rate in the evaluated baseline scenarios, with comprehensive accuracy scores of 0.931, 0.907, and 0.886, respectively. It outperforms conventional A*, RRT, and Hybrid A* algorithms in path length, smoothness, and safety. The method can support safe and efficient autonomous exploration by lunar rovers under the tested dynamic-illumination and multi-obstacle conditions and provides technical support for future deep-space exploration missions such as Chang’e and Tianwen.
To address the scarcity of high-precision control points on planetary surfaces and the accumulated drift of conventional relative-localization methods in deep-space exploration missions, this paper proposes a visual absolute-localization method based on salient-landmark contour matching and centroid-consistency constraints. Absolute localization is defined as estimating the rover position in the landing-site North-East-Down (NED) coordinate system or a map-projection coordinate system, rather than in image-pixel coordinates. Stable natural objects, including dunes and impact craters, are treated as generalized feature points. Local terrain is reconstructed from binocular navigation imagery; LiDAR is additionally used in the ground physical-equivalent experiment for multi-source terrain fusion. Multi-class cross-scale contour matching provides homologous landmark associations, after which centroid consistency aligns the local terrain with the global DOM/DEM reference frame. For ten Tianwen-1/Zhurong camera stations, the mean planar error was 0.458 m and the RMSE was 0.491 m. For five ground-test conditions, the mean planar error was 0.494 m and the RMSE was 0.526 m; all tested errors were below 1 m. Because the in-orbit reference DOM has a ground sampling distance of 1 m/pixel, the in-orbit sub-meter values indicate agreement with the adopted reference products and should not be interpreted as absolute accuracy independent of reference-map uncertainty. The results support the feasibility of natural-landmark-based map localization for future Chang’e and Tianwen missions.